An explainable hybrid stacked deep learning framework for forecasting PM10 concentrations in urban air

Accurate and explainable forecasting of particulate matter (PM10) is increasingly essential for managing urban air quality and protecting public health. This study proposed and evaluated a hybrid stacked deep learning architecture designed to enhance PM10 and urban air quality forecasting accuracy and to provide transparent explanations for its predictions. Using a self-designed neural network and Ridge regression (the meta-learner), PM10 prediction was accomplished based on LightGBM integration. Analysis was performed on the World Air Quality Index dataset, consisting of 1.8 million observations from 380 cities globally. To demonstrate its effectiveness, the hybrid model was benchmarked against traditional time series models (Autoregressive Integrated Moving Average [ARIMA] and Seasonal ARIMA) and machine learning models, including decision tree, extreme gradient boosting, random forest, and neural network, using the mean squared error (MSE), root MSE (RMSE), mean absolute error (MAE), and R2 metrics as evaluation metrics. Model explainability was accomplished using Shapley Additive Explanations and Local Interpretable Model-Agnostic Explanations analyses. The hybrid model achieved an R2 of 0.9916, MSE of 4.90, RMSE of 2.21, and MAE of 0.992, surpassing the other models’ performances and demonstrating strong reliability. The analysis determined the seven-day PM10 lag as the most important influential predictor, while other spatial parameters contributed minimally. The model’s ability to run efficiently on general-purpose computers further ensures accessibility for resource-constrained agencies. Overall, this study demonstrates the high predictive accuracy and interpretability of the proposed hybrid framework, offering a practical and informative tool for policymakers to improve air quality and public health outcomes.